WO2010101187A1 - Procédé et programme de création de base de données d'image, et procédé d'extraction d'image - Google Patents

Procédé et programme de création de base de données d'image, et procédé d'extraction d'image Download PDF

Info

Publication number
WO2010101187A1
WO2010101187A1 PCT/JP2010/053448 JP2010053448W WO2010101187A1 WO 2010101187 A1 WO2010101187 A1 WO 2010101187A1 JP 2010053448 W JP2010053448 W JP 2010053448W WO 2010101187 A1 WO2010101187 A1 WO 2010101187A1
Authority
WO
WIPO (PCT)
Prior art keywords
vector
image
feature
search
representative
Prior art date
Application number
PCT/JP2010/053448
Other languages
English (en)
Japanese (ja)
Inventor
貴行 本道
浩一 黄瀬
古橋　幸人
峯　泰治
Original Assignee
公立大学法人大阪府立大学
オリンパス株式会社
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 公立大学法人大阪府立大学, オリンパス株式会社 filed Critical 公立大学法人大阪府立大学
Priority to EP10748781.1A priority Critical patent/EP2405392B1/fr
Priority to US13/254,347 priority patent/US8649614B2/en
Priority to CN201080010386.4A priority patent/CN102341824B/zh
Priority to JP2011502784A priority patent/JP5527555B2/ja
Publication of WO2010101187A1 publication Critical patent/WO2010101187A1/fr
Priority to HK12105552.5A priority patent/HK1165067A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/50—Information retrieval; Database structures therefor; File system structures therefor of still image data
    • G06F16/56—Information retrieval; Database structures therefor; File system structures therefor of still image data having vectorial format
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
    • G06F16/33—Querying
    • G06F16/335—Filtering based on additional data, e.g. user or group profiles
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/40—Information retrieval; Database structures therefor; File system structures therefor of multimedia data, e.g. slideshows comprising image and additional audio data
    • G06F16/43—Querying
    • G06F16/432—Query formulation
    • G06F16/434—Query formulation using image data, e.g. images, photos, pictures taken by a user
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/50—Information retrieval; Database structures therefor; File system structures therefor of still image data
    • G06F16/53—Querying
    • G06F16/532—Query formulation, e.g. graphical querying
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/50—Information retrieval; Database structures therefor; File system structures therefor of still image data
    • G06F16/58—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually
    • G06F16/583—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually using metadata automatically derived from the content

Definitions

  • the present invention relates to an image database creation method, a creation program, and an image search method. More specifically, the present invention relates to a method for creating an image database used for specific object recognition using local features, a program for a computer to execute the creation method, and an image search method using the image database.
  • the specific object recognition is a process for determining which object in the image is exactly the same as the object in the other image.
  • image recognition Such processing can be used for detection of excess and deficiency of parts, detection of counterfeit products, replacement of bar codes, etc., and can be said to be highly practical.
  • the “object captured as an image” refers to an instance (search target) that is reflected in the image as a search question
  • a process for determining which object is exactly the same is a number of processes in advance. It can also be referred to as a process of searching for an image in which the same instance appears from an image database in which images are registered, that is, an image search process.
  • a method using a local feature is known.
  • identification is performed by expressing the image with local feature values extracted from the image by a predetermined procedure and comparing or collating with local feature values extracted from other images.
  • local feature amounts include SIFT (Scale-Invariant Feature Transform, for example, see Non-Patent Document 1) and PCA-SIFT (Principal Component Analysis-SIFT, for example, see Non-Patent Document 2). Since these local feature quantities are expressed as multidimensional vector quantities, they are also called feature vectors.
  • the number of local features extracted from a single image is usually about several thousand for VGA-sized images, and several tens of thousands for large numbers. Therefore, when the size of the recognition target image is large or many, the processing time required for collating those local features and the memory capacity required for storage become a problem.
  • Non-Patent Document 3 an approach of reducing the memory capacity necessary for recording individual local features has been proposed.
  • the amount of memory required to register each local feature amount in the image database is reduced by scalar quantization that reduces the number of bits of multi-value data representing the value of each dimension of the feature vector, and the entire image database
  • This method has an advantage that scalar quantization can be performed relatively easily by examining the distribution of values of each dimension of the feature vector in advance.
  • the concept of vector quantization has also been proposed.
  • D. Nister and others have proposed a method using a tree structure called Vocabulary Tree as one of vector quantization methods (see Non-Patent Document 4, for example).
  • this method in order to maintain a high recognition rate, the height of the tree structure must be increased, and there is a problem that the reduction effect cannot be sufficiently expected.
  • the present invention has been made in consideration of the above circumstances, and in the method of performing object recognition by neighborhood search using local features extracted from an image, the recognition rate of the object recognition is greatly reduced.
  • the present invention provides a method for reducing the storage capacity of an image database related to the object recognition and a program for causing a computer to execute the method. Further, the present invention provides a method for performing an image search using an image database created based on the above method.
  • the present invention corresponds to local features at different positions of a reference image to be matched with a search query image for object recognition, and represents the position and characteristics of each local feature as a vector position, vector length, and vector direction.
  • An extraction step of extracting a reference feature vector from the reference image a clustering step of creating a plurality of clusters of different reference feature vectors so that each reference vector belongs to one of them, and a reference feature vector of each cluster Selecting a representative vector of the cluster from the image, and registering the representative vector in an image database for object recognition in association with a reference image, wherein the clustering step has the same reference feature vectors at close vector positions.
  • Each cluster is created to belong to a cluster, and the selection process takes a long vector length reference.
  • the representative vector is selected by giving priority to a feature vector, and the search query image and the reference image generate at least one query feature vector from the search query image, and between the query feature vector and the representative vector
  • a method of creating an image database is provided in which a neighborhood search is applied and collated, and each process is executed by a computer.
  • the present invention corresponds to local features at different positions of the reference image to be collated with the search query image for object recognition, and the position and characteristics of each local feature are represented by a vector position, a vector An extraction step of extracting a reference feature vector represented as a length and a vector direction from the reference image, a clustering step of creating a plurality of clusters of different reference feature vectors so that each reference vector belongs to each of the clusters, and each cluster A selection step of selecting a representative vector of the cluster from the reference feature vector, and a step of associating the representative vector with a reference image and registering the representative vector in an image database for object recognition, and the clustering step includes: Reference feature vectors at close vector positions belong to the same cluster Each cluster is created, and the selecting step selects the representative vector in preference to a reference feature vector having a long vector length, and the search question image and the reference image are at least one query from the search question image.
  • the present invention provides at least one query feature representing a local feature from a search query image to be collated with a reference image registered in an image database for object recognition.
  • An extraction step for extracting a vector a matching step for matching by applying a neighborhood search between the query feature vector and the representative vector associated with each reference image, and the query feature vector being in the vicinity by the matching
  • the reference feature vector having a long vector length is preferentially selected from each cluster, and the image database stores the reference image and the representative vector extracted from the reference image in association with each other in advance.
  • An image search method in which each process is executed by a computer is provided. The procedure for generating the query feature vector from the search query image is the same as the procedure for extracting the reference feature vector.
  • each cluster is created so that reference feature vectors at close vector positions belong to the same cluster, and a predetermined number of representatives from each cluster is given priority over reference feature vectors having a long vector length. Since a vector is selected and matching is performed between the representative vector and the query feature vector, the memory capacity required for registering the feature vector in the image database can be saved as compared with the case where the representative vector is not selected. Can do. In addition, since each representative vector is registered from each cluster, that is, it is registered almost uniformly over the entire area without being biased to a part of the image, so the instances are unevenly distributed in the image or distortion due to geometric transformation. Even if it is received and received, it is possible to perform robust recognition.
  • the image database creation program according to the present invention has the same advantages as the image database creation method described above.
  • (d) is an example of an image obtained by capturing a part of a photograph of the instance. It is a graph which shows the result of the experiment example of this invention. The recognition rates for the search questions shown in FIGS. 3 (a), (b), (c), and (d) and their average recognition rates are shown.
  • the clustering step may generate a predetermined number of clusters. Even if instances are unevenly distributed in the image or are received due to distortion caused by geometric transformation, robust recognition is possible if the representative vectors are distributed almost uniformly over the entire area of the image. It can be carried out. The more vectors that are generated, the more uniformly the representative vectors are distributed. If the cluster granularity for which sufficiently robust recognition is performed is determined in advance experimentally, for example, and the clustering step generates a predetermined number of clusters, sufficiently robust recognition is possible. Can be realized.
  • one representative vector may be selected from each cluster.
  • feature vectors may be divided using a k-means method.
  • the feature vectors can be clustered so that they are evenly distributed over the entire area of the image.
  • One characteristic aspect of the present invention is that a reduction in the memory capacity of an image database used for image recognition is studied from the viewpoint of selection of local feature values, and a solution is provided. More specifically, the local feature amount is selected in consideration of the vector length (scale) of the feature vector and the uniformity of dispersion in the image space. According to the embodiments and experimental examples described below, even when using an image database in which the memory capacity is reduced to about 10% with respect to the memory capacity of the image database when the selection of local features is not performed, 98% A recognition rate could be obtained, demonstrating the effectiveness of the present invention.
  • the conventional memory capacity reduction method and image recognition processing by scalar quantization performed for specific object recognition will be described again.
  • the memory capacity reduction method by scalar quantization is a method for reducing the memory capacity of the image database by an approach different from the method of the present invention, and can be combined with the method of the present invention, and it is effective to combine them.
  • ⁇ Memory reduction method by scalar quantization
  • Non-Patent Document 3 proposes an approach called scalar quantization in order to reduce the memory capacity required for specific object recognition. This is to reduce the memory capacity by limiting the values that can be taken by each dimension of the feature vector representing individual local feature values to discrete values. That is, the value of each dimension is limited to a predetermined bit length. Although the number of local feature amounts to be registered in the image database is not changed, the memory capacity required for registering individual local feature amounts is reduced, so that the memory amount necessary for the image database as a whole is reduced.
  • each local feature is extracted from the reference image and the search query image by applying the PCA-SIFT technique.
  • Non-Patent Document 3 states that the recognition rate of image recognition hardly changes even if each dimension of the feature vector is expressed by 2 bits in the feature vector obtained by applying PCA-SIFT.
  • the value of each dimension of the feature vector extracted by PCA-SIFT is expressed by 16 bits when expressed by a short type integer. Therefore, if each dimension of the feature vector is scalar quantized and reduced to 2 bits, the feature vector alone has a memory capacity of about 1/8.
  • As an image database there is a necessary memory capacity in addition to storing feature vectors. However, it is stated that the memory capacity of the image database can be reduced to about 1/3 by taking this into consideration.
  • the image search is performed by collating the query feature vector with the reference feature vector.
  • the matching process calculates a distance between a query feature vector extracted from a search query image and a reference feature vector registered in an image database, and obtains a reference feature vector that is a neighborhood for each query feature vector. decide. Then, an image ID associated with the determined reference feature vector is obtained.
  • a process of determining the result of image recognition based on the result of collation is performed. In the process, voting is performed on the image ID for each query feature vector obtained by the matching process, and the reference image indicated by the image ID that has obtained the maximum vote is determined as a recognition result. As a result of scalar quantization, the accuracy of distance calculation is reduced. Still, the reason why the recognition rate hardly changes is that false image IDs are excluded thanks to the majority vote.
  • the local feature amount is extracted using the PCA-SIFT method.
  • the number of local feature amounts extracted from the reference image varies depending on the content of the reference image. All local feature values extracted from an image are registered in an unreduced image database in which selection of local feature values is not performed. For this reason, the number of registered local feature values differs greatly between different reference images. In a reference image from which a large number of local feature values are extracted, many similar local feature values may be extracted from a specific portion in the reference image. All similar local features need not be registered in the image database. Because it is similar, it is considered that it does not contribute much to the improvement of the recognition rate.
  • the maximum value of the number of local feature amounts extracted from one image into the image database is limited to R, thereby preventing an increase in memory capacity necessary for storing the reference feature vector. If the number of extracted reference feature vectors does not exceed R, all the extracted local feature amounts are registered in the image database. When the number of reference feature vectors exceeds R, a local feature value to be registered is selected based on the following idea.
  • a feature vector having a long vector length which is relatively resistant to changes in the shooting angle, is preferentially selected and registered in the image database. It can be said that the possibility that the entire search target is reflected in the reference image and the corresponding search question image to be the recognition result is not low. However, if a feature vector having a long vector length is unevenly distributed in a partial region of the reference image or the search question image, a portion other than the region becomes noise and it is difficult to search for a reference image corresponding to the search question. become. In order to deal with such uneven distribution of search targets, k-means clustering with the maximum number of clusters as R is performed on the coordinate value indicating the position of the reference feature vector in the reference image from which the reference feature vector has been extracted.
  • the reference feature vector in each cluster obtained by k-means clustering is preferentially selected from those having the largest vector length.
  • the selected reference feature vector is registered in the image database. That is, only representative vectors representing each class are registered in the image database. With this procedure, the reference feature vector is selected from the reference image substantially uniformly without deviation. Therefore, it is considered that the possibility of recognition can be increased even when only a part of the object to be searched is shown in the reference image.
  • ANN Approximate Nearest Neighbor, for example, see Non-Patent Document 5
  • ANN is a technique for performing an approximate nearest neighbor search at high speed using a tree structure. By performing the approximation, although the accuracy of vector matching is reduced, it is possible to reduce the processing time required for the search.
  • the image database used in the experiment was the one with 100,000 images registered as reference images.
  • the image database of 100,000 reference images is composed of three types of data sets A, B, and C.
  • A consists of 3,100 images collected using Google Image Search. Search keywords used to collect images are poster, magazine, cover, etc.
  • B consists of 18,500 images published on the PCA-SIFT site, and C is a photo sharing site. In flickr, it consists of 78,400 images collected by tags such as "animal", “birthday”, “food”, “japan”, etc. It mainly includes objects, nature photos, human photos, etc.
  • FIG. 2 shows an example of reference images collected by the above procedure.
  • centroid of the feature vector distributed in the divided feature space is obtained, and the feature vector in the space is replaced with the centroid vector.
  • a centroid vector is recorded, and vector quantization is performed by re-adding the image ID assigned to the replaced feature vector to the centroid vector.
  • This barycentric vector corresponds to a codeword of vector quantization and is often called a visual word.
  • Table 1 shows the number of local feature values registered in the image database of 100,000 reference images for each value of R.
  • FIG. 3 is an example of the obtained captured image.
  • the angle ⁇ of the optical axis of the camera with respect to the paper surface was changed to 90 °, 75 °, and 60 ° to obtain captured images.
  • a part of the paper was photographed at an angle of 90 °.
  • four captured images were obtained for each search target.
  • the captured image was reduced to 512 ⁇ 341 pixels to obtain a search query image, and a feature vector was obtained by PCA-SIFT. As a result, an average of 612 query feature vectors were obtained per search query image. [Determination of threshold value t]
  • A is a method of performing k-means clustering and selecting a feature vector having a long vector length among them.
  • B is a method of performing k-means clustering on an image space from each image and randomly selecting a local feature amount from the k-means clustering.
  • C is a method of selecting from each image in order from a feature vector having a long vector length.
  • D is a method of selecting a local feature amount randomly from each image.
  • the vertical axis represents the recognition rate
  • the horizontal axis represents the average recognition rate through the four data described below, with “average” at the left end.
  • “60 °” is the average recognition rate of search question images with a shooting angle of 60 °
  • “75 °” is the average recognition rate of search question images with a shooting angle of 75 °
  • “90 °” is the shooting angle of 90 °
  • the “partial” indicates the average recognition rate of the search question image obtained by photographing a part. From FIG. 4, when the whole image is shown, the method (A) has the best recognition rate.
  • Table 3 shows the recognition rate when the value of R is changed for method (A).
  • ⁇ indicates a case where the maximum number is not limited when the local feature amount is registered in the image database.
  • the present invention is an image database for a case where specific object recognition is performed on a large-scale image database such as tens of thousands or hundreds of thousands using local feature quantities such as SIFT (Scale-Invariant Feature Transform).
  • SIFT Scale-Invariant Feature Transform
  • p 1 , p 2 , p 3 , p 4 , p 5 , p 6 image feature vectors in the image database
  • q Search question feature vector
  • r Distance between vector p 1 and q, radius

Abstract

L'invention porte sur un procédé de création d'une base de données d'image comprenant une étape d'extraction consistant à extraire des vecteurs de caractéristiques de référence à partir d'une image de référence qui devra être comparée à une image d'interrogation d'extraction pour une reconnaissance d'objet, les vecteurs de caractéristiques de référence correspondant à des caractéristiques locales à différentes positions de l'image de référence et représentant la position et les caractéristiques de chacune des caractéristiques locales sous forme de position de vecteur, de longueur de vecteur et de direction de vecteur, une étape de regroupement consistant à créer une pluralité de groupes constitués de différents vecteurs de caractéristiques de référence, de telle sorte que chaque vecteur de référence appartient à l'un quelconque de la pluralité de groupes, une étape de sélection consistant à sélectionner le vecteur représentatif des groupes parmi les vecteurs de caractéristiques de référence de chacun des groupes, et une étape consistant à associer le vecteur représentatif avec l'image de référence et à enregistrer le vecteur représentatif associé à celle-ci dans la base de données d'image pour la reconnaissance d'objet, dans lequel, dans l'étape de regroupement, chacun des groupes est créé de telle sorte que les vecteurs de caractéristiques de référence à une position proche du vecteur appartiennent au même groupe, et dans l'étape de sélection, des vecteurs de caractéristiques de référence avec une longueur de vecteur longue se voient attribuer la priorité pour sélectionner le vecteur représentatif, et dans lequel l'image d'interrogation d'extraction et l'image de référence sont comparées entre elles par génération d'au moins un vecteur de caractéristiques d'interrogation à partir de l'image d'interrogation de l'extraction, et application d'une recherche locale entre le vecteur de caractéristiques d'interrogation et le vecteur représentatif, chacune des étapes étant exécutée par des ordinateurs.
PCT/JP2010/053448 2009-03-04 2010-03-03 Procédé et programme de création de base de données d'image, et procédé d'extraction d'image WO2010101187A1 (fr)

Priority Applications (5)

Application Number Priority Date Filing Date Title
EP10748781.1A EP2405392B1 (fr) 2009-03-04 2010-03-03 Procédé et programme de création de base de données d'image, et procédé d'extraction d'image
US13/254,347 US8649614B2 (en) 2009-03-04 2010-03-03 Method of compiling image database, image database compilation program, image retrieval method
CN201080010386.4A CN102341824B (zh) 2009-03-04 2010-03-03 图像数据库编辑方法、图像数据库编辑装置和图像检索方法
JP2011502784A JP5527555B2 (ja) 2009-03-04 2010-03-03 画像データベースの作成方法、作成プログラム及び画像検索方法
HK12105552.5A HK1165067A1 (zh) 2009-03-04 2012-06-07 圖像數據庫編輯方法、圖像數據庫編輯裝置和圖像檢索方法

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
JP2009050637 2009-03-04
JP2009-050637 2009-03-04

Publications (1)

Publication Number Publication Date
WO2010101187A1 true WO2010101187A1 (fr) 2010-09-10

Family

ID=42709742

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/JP2010/053448 WO2010101187A1 (fr) 2009-03-04 2010-03-03 Procédé et programme de création de base de données d'image, et procédé d'extraction d'image

Country Status (6)

Country Link
US (1) US8649614B2 (fr)
EP (1) EP2405392B1 (fr)
JP (1) JP5527555B2 (fr)
CN (1) CN102341824B (fr)
HK (1) HK1165067A1 (fr)
WO (1) WO2010101187A1 (fr)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2013105373A (ja) * 2011-11-15 2013-05-30 Yahoo Japan Corp データ取得装置、方法及びプログラム
JP2021068004A (ja) * 2019-10-18 2021-04-30 国立研究開発法人産業技術総合研究所 識別補助データ生成技術及び識別情報抽出技術

Families Citing this family (18)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103136228A (zh) 2011-11-25 2013-06-05 阿里巴巴集团控股有限公司 一种图片搜索方法以及图片搜索装置
US9734262B2 (en) * 2012-09-05 2017-08-15 Patrick DeLeFevre Method and system for understanding text
CN103324677B (zh) * 2013-05-24 2017-02-01 西安交通大学 一种可分级的快速图像gps位置估计方法
GB2516037A (en) 2013-07-08 2015-01-14 Univ Surrey Compact and robust signature for large scale visual search, retrieval and classification
WO2016043609A1 (fr) * 2014-09-18 2016-03-24 Empire Technology Development Llc Analyse sémantique latente en trois dimensions
CN105069457B (zh) * 2015-07-15 2020-02-11 杭州易现先进科技有限公司 图像识别方法和装置
CN106933867B (zh) 2015-12-30 2020-02-21 杭州华为企业通信技术有限公司 一种图像查询方法和装置
CN105718858B (zh) * 2016-01-13 2019-01-11 合肥工业大学 一种基于正负广义最大池化的行人识别方法
CN105718531B (zh) * 2016-01-14 2019-12-17 广州市万联信息科技有限公司 图像数据库的建立方法及图像识别方法
US10489712B2 (en) * 2016-02-26 2019-11-26 Oath Inc. Quality-based scoring and inhibiting of user-generated content
US10437878B2 (en) * 2016-12-28 2019-10-08 Shutterstock, Inc. Identification of a salient portion of an image
US10503775B1 (en) * 2016-12-28 2019-12-10 Shutterstock, Inc. Composition aware image querying
US11042586B2 (en) * 2016-12-29 2021-06-22 Shutterstock, Inc. Clustering search results based on image composition
US10248663B1 (en) * 2017-03-03 2019-04-02 Descartes Labs, Inc. Geo-visual search
CN108536769B (zh) * 2018-03-22 2023-01-03 深圳市安软慧视科技有限公司 图像分析方法、搜索方法及装置、计算机装置及存储介质
CN111177190B (zh) * 2018-11-13 2023-05-30 杭州海康威视数字技术股份有限公司 数据处理方法、装置、电子设备及可读存储介质
US11270155B2 (en) * 2019-11-26 2022-03-08 Dash Hudson Duplicate image detection based on image content
CN113159039A (zh) * 2021-02-09 2021-07-23 北京市商汤科技开发有限公司 图像识别方法及装置、电子设备和存储介质

Family Cites Families (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US6100901A (en) * 1998-06-22 2000-08-08 International Business Machines Corporation Method and apparatus for cluster exploration and visualization
JP3873793B2 (ja) 2002-03-29 2007-01-24 日本電気株式会社 顔メタデータ生成方法および顔メタデータ生成装置
JP4217664B2 (ja) * 2004-06-28 2009-02-04 キヤノン株式会社 画像処理方法、画像処理装置
US7596618B2 (en) * 2004-12-07 2009-09-29 Hewlett-Packard Development Company, L.P. Splitting a workload of a node
US8046363B2 (en) * 2006-04-13 2011-10-25 Lg Electronics Inc. System and method for clustering documents
JP5096776B2 (ja) * 2007-04-04 2012-12-12 キヤノン株式会社 画像処理装置及び画像検索方法
CN100587715C (zh) 2008-06-21 2010-02-03 华中科技大学 一种基于内容的鲁棒图像拷贝检测方法

Non-Patent Citations (7)

* Cited by examiner, † Cited by third party
Title
D. G. LOWE: "Distinctive image features from scale-invariant keypoints", INTERNAL JOURNAL OF COMPUTER VISION, vol. 60, no. 2, 2004, pages 91 - 110
D. NISTER, H. STEWENIUS: "Scalable Recognition with a Vocabulary Tree", PROC. CVPR 2006, 2006, pages 775 - 781
NOGUCHI, KISE, IWAMURA: "Experimental Study of Memory Reduction for Object Recognition based on Local Descriptors", COLLECTION OF PAPERS IN MEETING ON IMAGE RECOGNITION AND UNDERSTANDING (MIRU 2008, vol. 10-.3, 2008, pages 251 - 258
S. ARYA, D. MOUNT, R. SILVERMAN, A. Y. WU: "An optimal algorithm for approximate nearest neighbor searching", JOURNAL OF THE ACM, vol. 45, no. 6, pages 891 - 923
See also references of EP2405392A4 *
TAICHI JOTO ET AL.: "Bag-of-Keypoints Hyogen o Mochiita Web Gazo Bunrui", INFORMATION PROCESSING SOCIETY OF JAPAN KENKYU HOKOKU, vol. 2007, no. 42, 15 May 2007 (2007-05-15), pages 201 - 208, XP008167511 *
Y. KE, R. SUKTHANKAR: "PCA-SIFT: A more distinctive representation for local image descriptors", PROC. CVPR' 04, vol. 2, 2004, pages 506 - 513

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2013105373A (ja) * 2011-11-15 2013-05-30 Yahoo Japan Corp データ取得装置、方法及びプログラム
JP2021068004A (ja) * 2019-10-18 2021-04-30 国立研究開発法人産業技術総合研究所 識別補助データ生成技術及び識別情報抽出技術
JP7416400B2 (ja) 2019-10-18 2024-01-17 国立研究開発法人産業技術総合研究所 識別補助データ生成技術及び識別情報抽出技術

Also Published As

Publication number Publication date
JPWO2010101187A1 (ja) 2012-09-10
US8649614B2 (en) 2014-02-11
HK1165067A1 (zh) 2012-09-28
JP5527555B2 (ja) 2014-06-18
EP2405392A4 (fr) 2014-09-10
US20110317923A1 (en) 2011-12-29
CN102341824A (zh) 2012-02-01
EP2405392B1 (fr) 2015-08-05
CN102341824B (zh) 2016-05-18
EP2405392A1 (fr) 2012-01-11

Similar Documents

Publication Publication Date Title
JP5527555B2 (ja) 画像データベースの作成方法、作成プログラム及び画像検索方法
JP4883649B2 (ja) 画像認識方法、画像認識装置および画像認識プログラム
JP5294342B2 (ja) 物体認識用画像データベースの作成方法、処理装置および処理用プログラム
Amato et al. kNN based image classification relying on local feature similarity
Amato et al. Fast image classification for monument recognition
US8891908B2 (en) Semantic-aware co-indexing for near-duplicate image retrieval
Grycuk et al. Content-based image indexing by data clustering and inverse document frequency
Amato et al. Geometric consistency checks for kNN based image classification relying on local features
Ali et al. A leaf recognition approach to plant classification using machine learning
Khan et al. A hybrid CBIR system using novel local tetra angle patterns and color moment features
Paliwal et al. A score based indexing scheme for palmprint databases
Amato et al. Aggregating binary local descriptors for image retrieval
JP6017277B2 (ja) 特徴ベクトルの集合で表されるコンテンツ間の類似度を算出するプログラム、装置及び方法
JP5833499B2 (ja) 高次元の特徴ベクトル集合で表現されるコンテンツを高精度で検索する検索装置及びプログラム
Giouvanakis et al. Saliency map driven image retrieval combining the bag-of-words model and PLSA
Singh et al. Ensemble visual content based search and retrieval for natural scene images
Nguyen et al. A symbol spotting approach based on the vector model and a visual vocabulary
Guruprasad et al. Multimodal recognition framework: an accurate and powerful Nandinagari handwritten character recognition model
Dang et al. Srif: Scale and rotation invariant features for camera-based document image retrieval
Belhi et al. CNN Features vs Classical Features for Largescale Cultural Image Retrieval
Chen et al. Mobile visual search from dynamic image databases
Rajput Sketch based image retrieval using grid approach on large scale database
Amato et al. On knn classification and local feature based similarity functions
Gaillard et al. CNN features for reverse image search
He et al. Clustering-based descriptors for fingerprint indexing and fast retrieval

Legal Events

Date Code Title Description
WWE Wipo information: entry into national phase

Ref document number: 201080010386.4

Country of ref document: CN

121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 10748781

Country of ref document: EP

Kind code of ref document: A1

DPE1 Request for preliminary examination filed after expiration of 19th month from priority date (pct application filed from 20040101)
WWE Wipo information: entry into national phase

Ref document number: 2011502784

Country of ref document: JP

Ref document number: 13254347

Country of ref document: US

NENP Non-entry into the national phase

Ref country code: DE

WWE Wipo information: entry into national phase

Ref document number: 2010748781

Country of ref document: EP